ES-DETR: Real-time detection transformer with encover and soft-dropout.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Real-time object detection models like RT-DETR face limitations in capturing cross-channel feature interactions and robustness, especially for small or overlapping targets.
• Existing methods struggle with insufficient robustness and limited diversity in small-scale training datasets, hindering overall performance.

Purpose of the Study:

• To propose an improved real-time object detection model, the Enhanced Detection Transformer (ES-DETR), addressing the limitations of current methods.
• To enhance the model's ability to capture cross-channel feature interactions, improve robustness against information loss, and augment training data diversity.

Main Methods:

• Introduced 'Encover,' a spatial attention module replacing traditional flattened attention to learn global spatial features via cross-view image knowledge.
• Implemented 'Soft Dropout' (SD) to replace traditional dropout, suppressing features using a Gaussian distribution for stable, feature-wise dropout and increased detector robustness.
• Developed 'Grid Noise Augmentation' (GNA) by dividing images into grid patterns and applying Gaussian masks to mitigate real-world disturbances and improve training data diversity.

Main Results:

• ES-DETR demonstrates significant improvements in object detection tasks.
• The proposed methods (Encover, Soft Dropout, GNA) collectively enhance the model's speed, accuracy, and robustness.
• Experiments show ES-DETR outperforms existing models on several benchmark datasets.

Conclusions:

• ES-DETR effectively addresses the limitations of RT-DETR, offering superior performance in real-time object detection.
• The novel components of ES-DETR contribute to enhanced feature interaction, robustness, and generalization capabilities.
• The study provides a promising direction for developing more effective and robust real-time object detection systems.